Thanks very much for running the experiment, keishi@.
Here's my interpretation:
- the data suggests that heap compaction overhead on pause times isn't a major concern.
- that ~10 Blink pages (1 page = 128k) are on average released per compaction of the backing stores, which matches local data.
- that compacting the hash table and vector arenas reduce _total heap_ commit size by 7%. The data doesn't say when/how that differential was computed.
If the 7% hold up for a longer running renderer process, then compaction is arguably a worthwhile addition for keeping its Blink GC heap in a trim shape. Do keep in mind that compaction runs continually to combat fragmentation.
Comment 1 by sigbjo...@opera.com
, Dec 7 2016